PART I: ENCODING was an experiential happening in which ~60 audience members were guided to converse with someone they didn’t know. The conversation, and descriptions of the audience members, were then input as data into a custom LLM-based script-generating system.
DECODING was a live stage performance in three acts. As the audience took their seats, the collected data was used to generate a script. Using wearable displays, printers, and phones, actors received their lines in real time, performing a kind of embodied text-to-speech algorithm (which is maybe just a funny way of saying… they read the lines).
Different prompts to the system (additional moods and speaking styles) as well as choreography for the actors, layered upon each other until the performance became an abstract dance, ultimately descending to chaos.
The three acts were punctuated by my own real-time, semi-improvised, typed monologue, which served as the basis for a generative soundscape: an evolving musical score running throughout the piece.
My intention was, in many ways, experimental. I started with a question: what might be revealed by scrambling the machine learning pipeline, moving the "human-in-the-loop" from invisible roles (labelers, engineers, content creators) to explicit ones (via this audience-input, actor-output process)? Perhaps this scramble could help us capture something about this critical inflection point in technology, at a time when the relationship between machine learning and human knowledge isn’t (yet?) crystalized.
If you’re reading this in the future, I guess we’ll know a little more about whether that intention was met.
CREDITS:
Creative Technologist: Tommy Martinez
Systems Engineer: Yotam Mann
Production Assistant: Han Zhang
Performers: Payton Adamski, Zephyr Koczara, Olive Lafuente, Georgie McKeon, Eliane Mitchell, Arden Thomas, and Bryce Walsh
Additional Support from: Center for Performance Research, Onassis ONX, the Jerome Hill Foundation
photo credit: Elyse Mertz